Introduction
Nitrogen (N) is a critical nutrient for strawberry growth, directly influencing chlorophyll synthesis and overall fruit yield. Precise nitrogen management is essential for optimizing crop performance and minimizing environmental impact. While traditional laboratory techniques like Kjeldahl digestion provide accurate N measurements, they are destructive, time-consuming, and costly. Non-destructive alternatives, such as portable Near-Infrared (NIR) spectroscopy and SPAD meters, have emerged as promising tools. However, their accuracy is often compromised by "cultivar interference"—variations in leaf cell structure and light-scattering properties between different strawberry varieties. This study presents a "multi-cultivar" modeling approach designed to overcome these limitations.
Methodology
The researchers developed a hierarchical modeling framework that integrates cultivar recognition with variety-specific prediction models. The process follows four distinct stages:
- Spectral Acquisition: Reflectance data (350–2000 nm) were collected from 300 leaf samples across three cultivars: Hongyan (Hy), Ningyu (Ny), and Taoxun (Tx).
- Cultivar Recognition: To address cultivar interference, the team used the Random Frog algorithm to identify eight "fingerprint" wavelengths. A Random Forest (RF) classifier was then trained to identify the cultivar based on these specific spectral signatures.
- Optimal Wavelength Selection: For nitrogen prediction (measured via SPAD values), the Successive Projections Algorithm (SPA) was employed to isolate the most informative wavelengths for each cultivar, reducing data noise.
- Multi-Cultivar Pipeline: The final system operates as a smart switchboard. When a leaf is scanned, the RF classifier identifies the cultivar, and the system automatically applies the corresponding Partial Least Squares (PLS) regression model to estimate the SPAD value.
Results and Discussion
The study compared three modeling strategies: individual-cultivar models, a "hybrid" model (trained on a mix of all cultivars), and the proposed multi-cultivar model.
- Individual-Cultivar Models: These proved highly sensitive to variety; for instance, a model trained on Ningyu performed poorly when applied to Hongyan ($R_p = 0.147$).
- Hybrid-Cultivar Model: This offered better robustness than individual models but lacked precision because it essentially averaged out the unique optical characteristics of different leaves.
- Multi-Cultivar Model: This approach significantly outperformed both alternatives. It achieved an overall correlation coefficient of prediction ($R_p$) of 0.966 and a Root Mean Square Error of Prediction (RMSEP) of 0.468. The initial cultivar recognition step was highly successful, reaching an identification accuracy of 98.89%.
Conclusion
The hierarchical multi-cultivar approach effectively mitigates the impact of cultivar variation on NIR-based nitrogen assessment. This architecture is modular and "infinitely expandable," allowing for the addition of new strawberry varieties to the library without retraining the entire system. Future research should focus on testing the model under diverse environmental stresses and expanding the detection capabilities to other essential nutrients like phosphorus and potassium to provide a comprehensive plant health assessment.
Summary updated